miniCOIL: Sparse Neural Retrieval Done Right
summary
miniCOIL adds semantic weights to BM25 by reserving consecutive indices per token to store 4-dimensional meaning vectors and uses a one-layer dimensionality reducer on GINA embeddings, trained with triplet loss from 40 million sentences (about 8k sentences per word) to produce a 30,000-word vocabulary. The model is trained per-word (≈ one CPU each), fits classical inverted indices, adds negligible latency, and empirically ranks better than BM25 on out-of-domain benchmarks.
AI-written summary · about
chapters (9)
- 0:00Introduction to search
- 1:33The limits of current models
- 2:29Why typical approaches fail
- 7:29Developing a usable solution
- 10:51Architecting miniCOIL
- 16:17Model training and release
- 20:46Practical implementation
- 23:30Takeaways and future steps
- 25:55Audience Q&A
related talks
45mHybrid search: Greater than the sum of its parts?
16mMachine Learning on Source Code
46mNatural Language & Search, Improve your Search w/ Vector capabilities
30mSemantic Search and Neural Information Retrieval
32mBudget Friendly Semantic Search With Neural Sparse Search
41mHow to train your general purpose document retriever model
01h 28mBuilding a Semantic Search Engine
47mThe Neural Search Frontier #bbuzz